We built a streaming ML fraud platform for a card issuer that scores transactions in milliseconds, learns from analyst reviews and logs every decision. False positives fell 35%, and analyst workload fell 28%.
Industry: FinTech & Financial ServicesService: AI & Automation
Overview
Client
Fortitude Archives
Growing card issuer
Engagement snapshot
The challenge
The issuer's older fraud system used fixed rule chains that couldn't keep up with how customers' spending changed during seasonal peaks. Its false-positive rate was 35% higher than industry benchmarks.
Analysts were buried in alerts, legitimate customers were being declined, and the risk committee had already rejected ML models that couldn't explain their decisions.
The volume of false alerts built a 24-hour review backlog, which slowed the response to coordinated fraud waves.
Declining legitimate transactions cut into merchant lifetime value (MLV).
Without secure, traceable reason codes, the risk committee had rejected earlier ML models.
The solution
Behavioural features are computed on the transaction stream in under a second, with locked schemas that stop features drifting.
Review queues are split by risk band, and analysts' decisions feed back into the feature store to improve the model.
Every decision is saved with its contributing features and model version IDs, so it can be traced later.
Key features
The fraud model scores transactions using behavioural features computed on the live stream in under a second.
Business impact
12ms p95 scoring
Analyst decisions retrain the model, the likeliest fraud is reviewed first, and every decision carries reason codes.
Business impact
35% fewer false positives
Shared behavioural features update with every transaction, so the model always works from a consistent state.
Business impact
Zero feature drift
Each decision comes with its contributing weights as reason codes, to meet regulatory requirements.
Business impact
100% audit-ready decisions
Queues are ordered so analysts work on the most likely fraud waves first.
Business impact
28% load reduction
Before and after
Before
Static thresholds
Signals were judged in isolation, so normal changes in spending were often blocked.
After
Behavioural patterns
Unusual sequences and peer-group anomalies are spotted within sub-second streaming windows.
Before
Manual justification
There were no consistent reason codes, so audits relied on scattered analyst notes.
After
Audit-first decisions
Secure, traceable logs hold the exact feature state at the moment of each decision.
Before
Batch backlogs
Queues spiked at busy times, delaying manual transaction approvals.
After
Stream-native scale
Scoring stays under 12ms however much transaction volume spikes.
Results
35% Noise Reduction
Modelling each customer's behaviour in context cut false alerts while keeping fraud detection coverage.
Before: Rule-HeavyAfter: 35% Lower
28% Throughput Lift
Better routing let the existing team handle 28% more volume without adding headcount.
Before: CongestedAfter: Automated
12ms p95 Response Time
Streaming scoring takes 12ms at p95, so fraud checks don't slow down card authorisations.
Before: < 20msAfter: 12ms
Testimonial
“False alerts were taking up too much of our analysts' time. Since the new fraud model went live, false positives are down 35% and analyst workload 28%, and every decision carries a reason code we can trace.”
Joyce EdwardsHead of Fraud Operations, Fortitude ArchivesKey results
Trust and control
Each decision carries SHAP reason codes mapped to policy, so regulators can see why it was made.
COMPLIANCE READYUnneeded personal data is stripped out, and workflows are built to SOC 2 and GDPR requirements.
SOC 2-READYThe cloud-based streaming platform scales reliably during sudden traffic surges.
CLOUD NATIVEThe issuer owns 100% of the models, workflows and reusable modules from project completion.
100% OWNEDFaster delivery
Starting from tested components saved about 5.5 weeks of setup work. That time went into the active learning loop and the reason-code logs.
Our tested behavioural identity module gave the model a profile of each cardholder's usual activity to compare transactions against.
A reusable feature store module held the shared behavioural features and took analyst labels back in.
An existing monitoring platform tracked scoring times and model health in production.
Pre-built cost controls kept streaming compute spend predictable during seasonal traffic spikes.
We replaced fixed rule chains with streaming ML that scores each transaction against behavioural patterns in under a second. Analysts review cases in risk-banded queues, and their decisions feed back into the model. False positives fell by 35% while fraud detection coverage was maintained.
Analyst workload fell by 28%, so the existing team now handles 28% more volume without extra headcount. Before, false alerts built up a 24-hour review backlog. AI routing now puts the most likely fraud waves at the front of the queue, and scoring takes 12ms at p95.
Every decision is logged with SHAP reason codes, its contributing features and the model version used. The logs capture the exact feature state at the moment of the decision. That traceability mattered because the issuer's risk committee had rejected earlier ML models that could not explain themselves.
If rule-based checks are flooding your analysts and declining good customers, we can help you move to streaming ML that explains each decision.